2.7. Auditoría de Gestión
2.7.6. Construcción de indicadores
if one .of the frames Is a constituent of the other ('cos.istituel.nt' ere means rel'ated by the transitire closure of the 'ininediate constitlint' relation), it should trigger first.
If npeithler is a colnstituent, hope that some other lmchanisni, such as discourse context, will give one rule pweferencc. .
If each rule can be a constituent of the other,- the representation is probably. wrong anyway. I'm certainly not going to worry about it.
If. tIhe. triggerifng rtles interact so as to adhere to this principle, the effect
S w•t.ld be that the ftame- that triggers would be the 'lowest-level' conceptual phrase .(thinkitn of the frames in lerns of -a constituent hierarchy). There ýhould be no need to. allow a.. frame that contains only one constituent, since that. constituent would b.e a sinlle seiantic entity, and should be useful by itself.
An imnplementation problem is raised by this approach. Since we have been as..•tuiing thAat the .conceptual phra.ses are -parsed .into frames in a manner transpapen-t .to the othear tule.s (including the ones for parsing other conceptual phrassp), the only really conVenient trigger mechanism is attention-shJifting rules (.tvwhIch exjpliMjns
:the
use of attention-shifting rules. in- th.e current grammar).Unfaio'.tu.ately, the iinplrmoenltation of at tention-shifting rules in Parsifal constrains Ihe ni to ti'g11ge.r oil a single bilffer cell, in order to prevent violations of theoretical colnstra.lints on 'look.-aheadf' in the bilffer. It' would still be possible to implement t:-riggering by. having an attention-shifting rule match on the first cell of the
pattern for tfrigeing the frame, antd then ha.ve other. (norinal) rules -check the rest of the pattern agailnst the buffer, and restore it if none match. This addition to.the oifloplexity probably indicates a missed generalization, either in Parsifal or in our repr csc.Ji tati On.
Arlddiftonally3, from an ei•ginering point of view, It is not cqnvenient to have.
to examineo all the other rules in the system whenever a new conceptual frame is learned. Al this time, the probl-elp of how bes.t to specify the. triggering mechanism.
rmnains open.
6. ConiJlusiolnS
6.1 Where Ceonc•ptital. Phra;os Flt. III
One onf th.; major. linsnlved problems of both linguistics and artificial.
inliligrniuc is to define, the relationship between language and. knowledge. The solution, if there. is'one, w•ould have profound impact on both fields, and probably o.n many other fields as well.
This research represents an. attermpt Io.find a solution to that problem on a vrerry small scale, -- to develop a representation that captures the relationship lwPtwePen the synlax -ind the semantics of what I have termed 'conceptual phrases'.
The result is a spr'cification that consi.sts of a frame representation of an object, aiiigmoeni.er bIy a syntactic des!ription of the frame.
Coniceptual phrases fill. a niche between the various phases of the parsing process. "he different types of linguislic constructs, measured along .the axis of complexity in- emantic interpretation., can be seen as a spectrum. At one end of the ,Sp.ectrilln, .there are 'sentences', which are exceedingly gieneral conStructs in teims of th. me.anings, they can xppress. The users of the language pay a price for this gnep:erality -(as with most conipitiatlonal tasks), in terms of verbosity and the
Idifficuilty of senmintic interpretation.
At the other ,end. of the spectrum are individual words. Th'e process of issociating a wotd with its meaning is verry simple. Simply look it tup in the lexicon.. Wards, by th.emselves, have a very limited range of meaning, however.,**'
".'. whn .I ay "word". here, I mean an.Inrlldual lexlcal.lItem corresponding to a -single word in a sentence, wc.her.- all tlhr synfrctlc and discourse constrainil have already acled to select the correct word sense from a fixed list or alternatle'es
Coniceptual phrases are some'whern between these two ends of the spectrum.
They are more general .than words, but less general than clauses, for commlunicating concepts. Their range of neaninjg is sharply limited, but there are many more possibilities thai for single words., The process of seman tic interlltprltaliOn is much sinjmplr than for sentences, but more comnplex than for single words. It is interesting to speculate whether there are additional intermediate stages, each with its own special tradeoff between expressive power and efficiency.
Thor ere em Io be at least Iwo major schools of thought, among linguists and in. nrlificial intelli gence, about the relationship between syntax and s0mantics. One sct•oo( contends that the proces•es of language understanding and acquisition are mncrely facets of a geF neral ciognitive mechanism, and that syntax is an arbitrary set of rules for converting knovwlodge structures to linear streams of words, for the
puirpose of com.m.unicaing it. The other school believes that syntax and semantics ar ua parale (thuigh conneced lmechanisms), and that people have a special 'cognitive processor' or processors that is devotedr to the manipulation of language,
'Parsifal has been most strongly influenced by the second line of thought, yret the imechanisimn proposed herein is highly reminiscent of the first philosophy. The combination, of Parsifat's grammar languagle for higher-level constituents and this srpecification for flie.lo ] wer-level ones is a synthesis of the two viewpoints.
6.2
Di
ree1 Iions for Fut ure ]nsearchIn order to have a practical systeln for specifying conceptual phrases, there are- somie problemis which still must be dealt with. First, there is the problem discussred in Chapter 4 of ho-w to generate 'triggering' information from the aginented frame specification. This is part of a larger question about exactly how
34
the frafiies should initeract. Another task would be to determine !hnw much power is rnlquird if t:h syintacic dr-scr:iption, preferably stated in terms of linguiistic universalqs ralhFe"r than as ad-hoc observations about conceptual phrases in English.
'rI'l:e
are interesting linguifstic questions about conceptual phrases, if the analysis given. here is correct. The communications model developed in Chapter*2
seems to impliy .Iial a the procnrs .of generating a conceptual phrase is different from the pr;ocess of comnposi.ng a sentence. Can the representation given here be modified to make it hb-directional. (suitablte for the speaker as well. as. the listener)?.
Adrlitio•ially, there is Ihe question of how conceptual phrases are learned.
Since. -tli syntax. seemsi to be simpler than for higher. levels, are they easier -to learn? Chiap.ler . mentions. Ihal conceptual phrases can, be added to the language relalively freely. Does this imply that the process of learning conceptual phrases Is substant.ially different than lhe procefss of learning other syntactic rules? Tho prJiciples dteveloper. in Chatpter 5 could he implemented as an algorithm. which checks for potlential triggering. co••flicts with other rules. Does this imply that the deci~sion about how the frames should interact is made when the. rule is acquired?